[
https://issues.apache.org/jira/browse/AVRO-3408?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17515225#comment-17515225
]
Oscar Westra van Holthe - Kind commented on AVRO-3408:
------------------------------------------------------
IMHO, logical types are only related to schema evolution.
The way I understand reading Avro data is this:
* The write schema describes the bytes in a binary format
* The read schema describes the values to read from the data (with defaults if
data is missing)
* The reader object:
** reads the data described by the writer schema into the structure defined by
the read schema
(this part includes schema resolution / evolution, applying aliases, etc.)
** applies default values for missing data
** applies conversions for logical types
With this understanding, we now face a choice: should the reader apply logical
type conversions on the data as described by the write schema, or should the
logical type conversion be done on the data as placed in the read schema?
Option 1: the reader applies logical type conversions on the data as
transformed into the read schema (after schema resolution)
* this is the current situation
* logical type conversions only need to support the read schema
* schema resolution with logical types is very brittle:
** decimal type resolution only succeeds if precision & scale are identical
(as opposed to other numbers: int->long, int->float, long->double, etc. are all
supported)
** schema resolution for other logical types is not defined
Option 2: the reader applies logical type conversions on the data as described
by the write schema (before / instead of schema resolution)
* logical type conversions extend schema resolution
assuming conversions allow it, this enables (for example) a widening conversion
int->decimal(10,0), or decimal(3,0) -> decimal(6,2)
* logical type conversions may need to support the write schema as well (if
there's no logical type on the write schema)
* schema resolution depends on presence of logical types and conversions
> Schema evolution with logical types
> ------------------------------------
>
> Key: AVRO-3408
> URL: https://issues.apache.org/jira/browse/AVRO-3408
> Project: Apache Avro
> Issue Type: Improvement
> Components: java
> Affects Versions: 1.11.0
> Reporter: Ivan Zemlyanskiy
> Assignee: Ivan Zemlyanskiy
> Priority: Major
> Labels: pull-request-available
> Fix For: 1.12.0
>
> Time Spent: 4h 10m
> Remaining Estimate: 0h
>
> Hello!
> First of all, thank you for this project. I love Avro encoding from both
> technology and code culture points of view. (y)
> I know you recommend migrating schema by adding a new field and removing the
> old one in the future, but please-please-please consider my case as well.
> In my company, we have some DTOs, and it's about 200+ fields in total that we
> encode with Avro and send over the network. About a third of them have type
> `java.math.BigDecimal`. At some point, we discovered we send them with a
> schema like
> {code:json}
> {
> "name":"performancePrice",
> "type":{
> "type":"string",
> "java-class":"java.math.BigDecimal"
> }
> }
> {code}
> That's a kind of disaster for us cos we have pretty much a high load with ~2
> million RPS.
> So we start to think about migrating to something lighter than strings (no
> blame for choosing it as a default, I know BigDecimal has a lot of pitfalls,
> and string is the easiest way for encoding/decoding).
> It was fine to make a standard precision for all such fields, so we found
> `Conversions.DecimalConversion` and decided at the end of the day we were
> going to use this logical type with a recommended schema like
> {code:java}
> @Override
> public Schema getRecommendedSchema() {
> Schema schema = Schema.create(Schema.Type.BYTES);
> LogicalTypes.Decimal decimalType =
> LogicalTypes.decimal(MathContext.DECIMAL32.getPrecision(),
> DecimalUtils.MONEY_ROUNDING_SCALE);
> decimalType.addToSchema(schema);
> return schema;
> }
> {code}
> (we use `org.apache.avro.reflect.ReflectData`)
> It all looks good and promising, but the question is how to migrate to such
> schema?
> As I said, we have a lot of such fields, and migrating all of them with
> duplication fields with future removal might be painful and would cost us a
> considerable overhead.
> I made some tests and found out if two applications register the same
> `BigDecimalConversion` but for one application the `getRecommendedSchema()`
> is like the method above and for another application the
> `getRecommendedSchema()` is
> {code:java}
> @Override
> public Schema getRecommendedSchema() {
> Schema schema = Schema.create(Schema.Type.STRING);
> schema.addProp(SpecificData.CLASS_PROP, BigDecimal.class.getName());
> return schema;
> }
> {code}
> so they can easily read each other messages using _SERVER_ schema.
> So, I made two applications and wired them up with `ProtocolRepository`,
> `ReflectResponder` and all that stuff, I found out it doesn't work. Because
> `org.apache.avro.io.ResolvingDecoder` totally ignores logical types for some
> reason.
> So as a result, one application specifically told "I encode this field as a
> byte array which supposed to be a logical type 'decimal' with precision N",
> but another application just tries to convert those bytes to a string and
> make a BigDecimal based on the result string. As a result, we got
> {code:java}
> java.lang.NumberFormatException: Character ' is neither a decimal digit
> number, decimal point, nor "e" notation exponential mark.
> {code}
> In my humble opinion, `org.apache.avro.io.ResolvingDecoder` should respect
> logical types in _SERVER_ (_ACTUAL_) schema and use a corresponding
> conversion instance for reading values. In my example, I'd say it might be
> {code}
> ResolvingDecoder#readString() -> read the actual logical type -> find
> BigDecimalConversion instance ->
> conversion.fromBytes(readValueWithActualSchema()) ->
> conversion.toCharSequence(readValueWithConversion)
> {code}
> I'd love to read your opinion on all of that.
> Thank you in advance for your time, and sorry for the long issue description.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
